Skip to content

audit(#419): declared-only rescore of referee_deployable and plausible_distractor - #429

Open
maximinl wants to merge 1 commit into
mainfrom
audit/419-declared-only
Open

maximinl wants to merge 1 commit into
mainfrom
audit/419-declared-only

Conversation

@maximinl

Copy link
Copy Markdown
Collaborator

Summary

  • Offline declared-only audits for MedQA referee_deployable and plausible_distractor (Rescore referee arms from committed caches using declared-only accounting #419), replaying committed caches with zero API calls.
  • New scripts write *_declared_audit.json only; historical *_summary.json / .jsonl are untouched.
  • Deployable: 39/40 cases fully declared; precision 0.682 → 0.714, clean FP 7 → 6 — headline unchanged.
  • Plausible distractor: 31/103 fully declared; adoption gap 0.146 → 0.226, p 0.0059 → 0.039 — still significant at α=0.05, but most cases had at least one undeclared draw (ranking prompts especially).

Notes

  • MedMCQA mirrors are ready to run the same scripts once a local manifest is staged; not included here (no MedMCQA raw on this machine).
  • MIMIC referee_deployable remains blocked (call cache gitignored).

Test plan

  • python -m experiments.referee.declared_only_audit_deployable --manifest <medqa_manifest.csv> (new_api_calls_this_run=0, replay matches committed)
  • python -m experiments.medqa.declared_only_audit_plausible_distractor --manifest <medqa_manifest.csv> (same)

Closes #419

Made with Cursor

…ctor

Offline cache replays report current vs declared_mcq_choice-eligible metrics
without touching historical summaries. MedQA deployable is stable; plausible
distractor drops to 31/103 fully declared cases and widens the adoption gap.

Co-authored-by: Cursor <cursoragent@cursor.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

Rescore referee arms from committed caches using declared-only accounting

1 participant